Earticle

The present study aims to compare the translation quality of human translation and AI translation-specifically neural machine translation (NMT) and large language model (LLM)-based translation—in the context of Korean–Japanese translation. The source texts comprised informative texts (news articles) and expressive texts (columns), while the target texts included professional human translations as well as AI translations generated by Google, Papago, DeepL, and ChatGPT in 2023 and 2025. Translation quality was assessed using BLEU and TER metrics computed with SacreBLEU to ensure reproducibility, and complemented by a qualitative analysis employing Molina & Hurtado Albir’s(2002)translation techniques at the sentence level. The results revealed that DeepL achieved the highest BLEU and TER scores, whereas ChatGPT employed a broader range of techniques and produced translations most comparable to human translations. These findings highlight the limitations of automatic metrics and demonstrate the importance of combining quantitative metrics with qualitative and human evaluation to achieve a more comprehensive understanding of AI translation performance.

본 연구는 중국어 ‘黑+중첩접미사’를 포함하는 색채어 형용사 예문의 인공신경망 기계 번역 과 대화형 AI 기계 번역의 결괏값을 비교 분석하는 것을 목적으로 한다. 먼저, 중국어의 경우 국내와 국외의 대화형 AI 언어모델 중 더 높은 품질의 번역을 제공하는 것은 국내 네이버사의 ‘Hyper Clova X’이다. 다음으로, ‘黑+중첩접미사’류 형용사를 포함한 구문의 기존 신경망 기 계번역 프로그램 ‘Papago’와 대화형 AI 언어모델 ‘Hyper Clova X’에서 번역 결과 대체로 Clova X의 번역 품질이 더 높았다. 끝으로, 중한번역에서 대화형 AI 언어모델 활용 시 사용자 가 고려해야 하는 부분으로는 한국어 조사, 어미 등에 오류 여부를 확인하고, 입력한 문장이 축약된 정보를 포함하고 있을 때 출력값에 오류가 있을 가능성이 크므로 주의해야 한다. 본 연구를 토대로 더욱 다양한 중국어 결합구조를 활용한 텍스트를 통하여 여러 상황에 AI 기술과 인간이 함께 공존하고 발전해나갈 수 있도록 관련 분야의 연구가 지속되기를 기대한다.

This study aims to compare and analyze the results of neural network machine translation and generative AI machine translation for color adjectives containing the Chinese '黑+compound suffix' Firstly, within the realm of Chinese language, 'Hyper Clova X' by the domestic company Naver provides higher-quality translations among both domestic and international conversational AI language models. Subsequently, when examining translations of sentences with '黑+compound suffix' type adjectives, the existing neural network machine translation program 'Papago' and the conversational AI language model 'Hyper Clova X' were compared, revealing that, in general, Clova X exhibited superior translation quality. Finally, in utilizing conversational AI language models for Chinese translation, users should be attentive to potential errors in Korean particles, verb endings, and other grammatical elements. Additionally, caution is advised when the input sentence contains condensed information, as this may increase the likelihood of errors in the output. Building upon this research, it is anticipated that ongoing studies in related fields will continue to explore diverse Chinese syntactic structures, facilitating the coexistence and development of AI technology and human communication in various contexts.

본 연구는 감정 중심 서사인 손원평의 『아몬드』를 분석 대상으로 삼아 한국어 원문과 일본어 번역문(인간 번역 및 AI 번역: GPT–5, Gemini 2.5 Pro)간의 감정 재현 양상과 한계를 검토하였다. 분석은 다국어 감정 분석 모델을 활용해 문단 단위의 감정 점수 (Polarity)를 산출하는 정량 분석과, 수치만으로 설명되지 않는 표현상의 차이를 검토하는 정성 분석을 병행하였다. 분석 결과, AI 번역은 원문 기반임에도 불구하고, 결과적으로 인간 번역과 매우 높은 감정 점수 상관성을 보였다. 이는 생성형 AI가 문학 텍스트의 정서적 흐름을 일정 수준 재현할 수 있으며, 인간 번역의 정서적 구성과 일부 지점에서 유사한 패턴을 만들어 낸다는 점을 시사한다. 그러나 『아몬드』는 감정을 직접적으로 표출하기보다, 상황적 단 서, 서술 리듬 등 간접적 장치를 통해 정서를 형성하는 특성을 지니고 있어, 감정 분석 모델에서는 중립적이거나 약화된 정서로 처리되는 경향이 있었다. 또한, 아이러니, 혼합 정서, 정서적 거리두기 등 문학 텍스트 특유의 장치는 단순한 Polarity 수치로 환원하기 어려웠다. 결론적으로, 감정 점수는 문학 번역에서 정서적 재현의 경향성과 차이를 비교하는 유용한 보조 지표로 기능할 수 있지만, 문학적 감정의 복잡성을 온전히 설명하기 위해서 는 정성적 해석과의 결합이 필수적이다. 이러한 복합적 접근은 감정 점수의 해석 범위를 넓히고, 번역학･감정 계산학･디지털 인문학을 연결하는 새로운 분석 틀을 마련한다는 점에서 의의를 지닌다.

This comparative study utilizes human and AI (GPT-5, Gemini 2.5 Pro) translations to examine the reproduction and limitations of emotional expression between the Japanese and Korean translations of Son Won-pyung's Almond. Using a multilingual emotion analysis model, paragraph-level polarity scores were calculated and complemented by qualitative analysis of stylistic differences. Despite being source-based, results show that AI translations showed notable alignment with the human translation in emotional scores, suggesting a partial consistency in emotional patterns rather than a complete capture of literary emotional flow. Simultaneously, specific limitations of this approach were revealed by differences in how emotional cues are realized in Korean and Japanese, model-level biases, and literary devices such as irony or emotional distancing. These findings suggest that, when combined with qualitative interpretation, polarity scores can serve as a useful auxiliary metric for assessing emotional reproduction in translation. The study indicates the possibility of an analytical framework connecting translation studies, affective computing, and digital humanities.

This study examines how collaborative translation has been approached in Korean scholarship and explores its conceptual framework and potential for development. Unlike in international research, where the term is more clearly defined, collaborative translation remains underused and insufficiently theorized in Korea. To address this gap, the study analyzes 254 academic papers to identify prevailing trends and limitations. Findings show that machine-centered approaches dominate (61.4%), focusing mainly on MT evaluation and translation education, while studies on professional collaboration between human translators and MT systems are relatively scarce. Human-centered studies (38%) encompass both multi-translator models—such as online collaboration, relay translation, and co-translation—and single-translator models involving collaboration with editors, proofreaders, clients, and others. Despite the widespread practice of co-translation, it remains underexplored in academic literature, exposing a gap between industry and academia. Existing classification models also fail to account for cases where a single translator assumes multiple roles, underscoring the need for a more nuanced framework that considers the intensity, stages, and dynamics of collaboration. This study highlights the need for a clearer and more comprehensive conceptualization of collaborative translation, and calls for future research that bridges theoretical insights with evolving technological and professional practices.

This study examines the performance of four major AI translation tools Google Translate, Papago, DeepL, and ChatGPT in translating poetry between Korean and English in both directions. Poetry translation presents unique challenges, including cultural context, stylistic intricacies, and emotional resonance, which machine translation tools traditionally struggle to capture. The research utilizes Juliane House’s 2015 Translation Quality Assessment model(the TQA), combined with a custom numerical rubric, to evaluate both covert and overt errors in translation. To ensure cultural and poetic variety, the study analyzes two canonical poems: Seo, Jeong-ju’s “Beside the Chrysanthemums” (Korean) and Robert Frost’s “The Road Not Taken” (English). These serve as case studies for both Korean-to-English (K2E) and English-to-Korean (E2K) translation tasks. Evaluation scores from both human and AI assessors reveal that ChatGPT achieved the highest performance, scoring 93/100 on K2E, closely rivaling human translations, while DeepL scored significantly lower at 72/100. Results also indicate that translations from English to Korean (E2K) were consistently more challenging, with average scores dropping by 7.75 points across all tools. Strong correlations between AI and human evaluation scores (r 0.83) suggest that AI models, particularly ChatGPT, can support high-quality literary translation assessments. The study highlights the emerging role of generative AI as an assistant not a replacement in poetry translation workflows, suggesting a promising future for AI-human collaboration in preserving literary nuance and emotional depth.

The problem of discrimination due to information alienation among Korean sign language users is continuously mentioned, and sign language translation research is actively being conducted to solve this problem. However, due to technological limits in translating text into Korean Sign Language, the need for specialist equipment causes annoyance and spatial constraints. Furthermore, it isn't easy to replicate the vocabulary and grammatical structure of the Korean language in Korean Sign Language. Furthermore, the service's commercialization is complicated by the need for more nonmanual signal (NMS) identification technology.

The purpose of this research was to investigate the characteristics of Korean college students’ writings, which have been produced without or with the help of machine translation tool in the classroom. Specifically, this research attempted to investigate the linguistic characteristics of the students’ writings, and types of errors identified in the writings. Twelve pieces of writings from three college students were collected for analysis. Two online word analysis programs, Word Counter (2023) and LIWC-22 (2023), were employed for data analysis. The findings of data analysis found out that 1) The students’ drafts consisted of 22.8 sentences including 303.9 words in 3.6 paragraphs on average. 2) In the students’ drafts, ‘unique’ words (46.8%) were included a lot more than ‘difficult’ words (27%), and students tended to write their essay writings in an unfiltered or impromptu way rather than an analytical way regardless of their English language proficiency levels. 3) The highest frequency of errors was seen in grammatical errors (41.7%) followed by lexical errors (31.6%). Based on the research findings, pedagogical implications and suggestions for the effective use of machine translation in English writing classes were presented.

This study investigates the syntactic complexity in English prose between the writings of Chinese scholars and the corresponding translations generated by AI-based machine translation systems. A corpus of 100 English abstracts written by Chinese scholars and 300 English abstracts translated by ChatGPT 4.0, Google Bard and Microsoft Bing was constructed. These texts were analysed using 14 measures of syntactic complexity as defined by the L2 Syntactic Complexity Analyzer (Lu, 2010). The analysis revealed that when comparing the original Chinese-English texts with the outputs of machine translation systems, significant differences were found in 13 of the 14 syntactic measures. Conversely, when comparing the translations from ChatGPT 4.0, Bard and Bing, significant differences were found in 10 of the 14 measures. This research advances the understanding of machine translation systems and has relevant implications for pedagogy and assessment in the field.

Driven by evolvement of AI technology, a series of innovative technologies and tools, such as crowdsourcing translation, remote-video interpreting, human-computer interactive translation, VR/AR interpreting, smart translation pen, portable translation device, and AI simultaneous interpreting technology have made their presence and practical application in various fields and scenes. Technology is reconstructing the traditional translation in an unprecedented way, leading to heated discussion on reform of translation profession and translation education. As translation technology is an undeniable trend, it is necessary to redefine translation competence saying that technical competence is a must for future translators. Questionnaires and interviews were conducted via telephones and emails on the current situation of translation technology curriculum in MA/MTI Translation/Interpreting programs in China, Japan, South Korea, the United States, the United Kingdom, France, Germany, Spain, Switzerland, Australia, etc. This study will present the common problems found in those programs and provide constructive suggestions to them, as maybe referred by future translation technology education in a global view.

본 연구는 인공지능 다언어 번역 환경의 확산이 외국어 텍스트 이해 방식에 변 화를 가져오고 있다는 점에 주목하여 다국적 교실에서 공적 담화 번역 활동을 활 용한 한국어 수업 모형을 제안하는 것을 목적으로 한다. 최근 신경망 기반 번역 기술의 발전으로 학습자는 다양한 언어 번역 결과에 쉽게 접근할 수 있게 되었으나 교육 현장에서는 인공지능 번역이 주로 이해 보조 도구로 활용되는 경향이 있 었다. 이에 본 연구는 인공지능 번역을 정답 제공 수단이 아니라 비교와 분석을 촉진하는 학습 자원으로 재개념화하였다. 본 연구에서는 경제 기사를 공적 담화 텍스트로 선정하고 인공지능 다언어 번역 자료를 기반으로 언어권별 협력 번역, 조별 발표 및 토론, 담화 재해석과 메타 인식 단계로 구성된 수업 모형을 설계하였 다. 경제 기사 텍스트는 인과 관계와 정보 위계가 밀집된 담화 구조를 지니며 번역 활동과 결합될 때 학습자가 의미 관계를 분석하고 재구성하는 과정을 촉진한다. 제안된 모형에서 인공지능 번역 결과는 비교 대상 자료로 활용되며 학습자는 다양 한 해석 가능성을 검토하고 번역 선택의 근거를 형성하게 된다. 본 연구는 인공지 능 번역 활용 수업이 읽기, 번역, 쓰기 활동을 담화 재구성 과정 속에서 통합적으로 운영될 수 있음을 제시하고 다국적 교실의 언어적 다양성을 학습 자원으로 전환하 는 수업 설계의 가능성을 제안한다.

This study proposes an instructional model for Korean language education that integrates multilingual AI translation with public discourse translation activities in multinational classrooms. With the rapid development of neural machine translation technologies, AI translation tools have increasingly influenced how learners access and interpret foreign language texts. Rather than treating AI translation as a tool for providing correct answers, this study reconceptualizes it as a pedagogical resource that promotes comparative analysis and discourse reconstruction. Focusing on economic news articles as representative public discourse texts, the proposed model combines discourse analysis, language-group collaborative translation, and presentation-based reflection activities. Economic news texts contain dense causal relations, information hierarchies, nominalized expressions, and connective structures, which require learners to interpret meaning beyond sentence-level comprehension. By engaging learners in multilingual translation comparison and collaborative rewriting tasks, the model encourages learners to analyze discourse structure, negotiate meaning, and reconstruct texts in Korean. The instructional design consists of three stages: discourse awareness, collaborative meaning negotiation through language-group translation, and presentation-based reflection leading to metalinguistic awareness. In this process, AI-generated multilingual translations function as comparative materials rather than authoritative outputs, positioning learners as active decision-makers in meaning construction. The study discusses the educational implications of integrating AI translation into Korean language instruction and suggests that AI-supported learning environments can facilitate discourse-level comprehension and collaborative knowledge construction among intermediate and advanced learners. Although the proposed model has not yet been implemented in an actual classroom setting, it offers a pedagogical framework for rethinking the role of AI translation in Korean language education and for designing discourse-centered instruction in multilingual learning environments.

This study investigates how effectively generative AI models reproduce the stylistic features of Korean emotional essays in Japanese translation. Using I Decided to Live as Myself, we compared human translation with three AI-generated translations. Semantic similarity was measured with BERTScore, and stylistic variation was assessed through sentence-level embedding distances based on SentenceTransformer. The quantitative results show stable semantic fidelity across AI models but considerable divergence in stylistic embeddings, indicating stylistic variation despite preserved meaning. Qualitative analysis further revealed differences in sentence structuring, emotional tone, cultural contextualization, and information organization between human and AI translations. The findings demonstrate that meaning and style can diverge substantially in emotional-text translation and that AI models reproduce stylistic cues only partially. This study provides foundational insight for developing stylistic evaluation metrics and advancing research on AI translation of emotion-centered texts.

This study compares how Papago, Google Translate, and ChatGPT perform when translating police dialogues, using BLEU scores, manual assessment, and error analysis. The analysis results are as follows. All three tools showed low BLEU scores compared to reference translations. In manual evaluation, ChatGPT significantly outperformed others, scoring 8.6 out of 10, compared to Google Translate's 6.5 and Papago's 5.6. Error analysis confirmed ChatGPT's superiority, with only 41 errors, while Google Translate and Papago produced 123 and 152 errors respectively. Across all tools, substitution was the most common accuracy error, followed by omission and addition. These results suggest that ChatGPT is the most reliable tool for police communication with foreign nationals.

AI도구의 발전은 일본어를 비롯한 언어 학습에도 큰 변화를 가져오고 있다. 본 연구에서는 AI시대의 일본어 연구 성과를 바탕으로 일본어교육의 실천 사례를 소개하였다. 먼저, AI시대의 일본어 연구에서는 텍스트 마이닝 기법을 활용하여 AI 관련 일본어 연 구의 성과와 동향을 기술하였다. 그 방법으로 Google Scholar 검색 도구를 사용하여 “AI”와 “日本語”를 모두 포함하는 발표･연구 논문에 수록된 연구타이틀을 추출하였다. 이를 통해 얻은 176건의 데이터를 대상으로 텍스트 마이닝을 실시하였다. 추출어 분석 결과, AI 관련 일본어 연구는 ‘일본어교육’ 분야에서 가장 활발히 진행되고 있음을 확인 하였다. 구체적인 연구 대상으로는 [翻訳]이 가장 많았으나, 그밖에도 [作文] [文章] [音 声] [文法] 등 폭넓게 AI가 활용되고 있다. 다음으로, AI번역 도구를 활용한 일본어교육 실천 사례를 제시하였다. 한국의 대학 에서 진행한 “이문화 커뮤니케이션” 수업의 활동으로서 ‘AI 도구를 활용한 동영상 자막 제작’을 소개하였다. 동영상 자막 제작 활동의 의의로는 동영상 콘텐츠를 활용함으로써 수업에 대한 학생들의 관심을 끌 수 있었다는 점이다. 또한, 학생들이 직접 동영상을 선택하고 스스로 번역과 자막을 만들어 발표함으로써 수업에 적극적으로 참여하는 모 습을 보였다. 전체적인 성과로는 교사의 언어 능력과 학생의 디지털 능력을 융합한 형 태의 교실활동으로서 AI 번역을 통한 언어 능력 및 디지털 활용 능력의 교육적 가능성을 확인하였다.

There are significant changes to learning languages, such as Japanese, with the advancement of AI tools. In the AI era, text mining techniques are employed to describe the achievements and trends in AI-related Japanese language research. Specifically, I used the Google Scholar search tool to detect research titles containing both “AI” and “日本語” (Japanese language) from published papers and research articles. The 176 data points obtained were then mined for text. Analysis of the extracted terms confirmed that AI-related Japanese language research is most actively conducted in the ‘Japanese language education’ field. The most common specific research subject was [translation], but AI is also being widely used in areas, such as [Composition], [Sentence], [Phonetics], and [Grammar]. In this study, I present practical examples of Japanese language education using AI translation tools. My study introduces ‘Video Subtitle Creation Using AI Tools,’ an activity implemented in an ‘Inter-Cultural Communication’ course at a Korean university. The significance of the Video Creation (translation and subtitling) activity was that it captured the interest of students in the class by utilizing video content. Furthermore, students actively participated by selecting the videos themselves, independently created translations and subtitles, and presented their work. Overall, this activity was confirmed to be effective to enhance both language proficiency and digital literacy skills through AI translation. This represents a classroom activity that combined the linguistic competence of the teacher with the digital capabilities of the students. Overall, the findings of my study confirm the educational potential of classroom activities that integrate linguistic proficiency of teachers with digital literacy of students by utilizing AI translation to enhance both language skills and digital application abilities.

This study analyzes the evolution of AI translation technology from the perspective of the philosophy of language. AI translation has advanced to a stage where it infers context and meaning through NMT. This evolution has triggered a philosophical debate and fear regarding whether AI can encroach on the domain of pragmatics, including human intention and emotion. This debate is broadly divided into two main viewpoints: anti-AI and pro-AI. Meanwhile, Quine and Wittgenstein view meaning not as a fixed entity, but as something determined by usage, context, and communal consensus. This perspective is analogous to AI, however, and AI lacks deep consensus and social practice of a human community, thus showing limitations in translating pragmatic discourse that goes beyond the straightforward narrative level. Beyond these technical limitations, a critical challenge is that AI translation faces major hurdles: algorithmic bias, which amplifies social prejudices inherent in the training data, and linguistic homogenization, which threatens the diversity of minority languages. As a methodical solution, this article proposes defining AI translation as a distinct 'sociolect' (or 'social language') with its own unique rules and characteristics.